nllb-sampled-500k

This dataset contains sampled bilingual sentence pairs derived from allenai/nllb. It is published in a convenient Hugging Face layout with one subset per language pair, for example ace_Latn-ban_Latn.

The dataset is intended for multilingual representation learning, translation-alignment training, cross-lingual retrieval experiments, and other tasks that benefit from broad bilingual positive pairs.

Source and transformation

• Original source: allenai/nllb
• Upstream source families: NLLB bitext sources, including CCMatrix-derived data where applicable, as documented by the NLLB dataset card
• Transformation in this release: each language-pair subset is sampled to at most roughly 500k rows and exported as Parquet for direct loading with the datasets library

No semantic relabeling is applied: each row remains a bilingual sentence pair from the upstream source.

Data fields

• source: text in the first language of the subset name
• translated_text: paired text in the second language of the subset name

For a subset named src-tgt, source corresponds to src and translated_text corresponds to tgt.

Loading

from datasets import load_dataset

ds = load_dataset("hotchpotch/nllb-sampled-500k", "eng_Latn-jpn_Jpan", split="train")
print(ds.column_names)  # ['source', 'translated_text']

License

This dataset follows the license and usage terms of the original source data. The immediate upstream dataset is allenai/nllb, whose dataset card declares the NLLB license as ODC-BY. Some upstream bitext sources are CCMatrix/Common Crawl derived; users who need strict compliance should review the original NLLB and source-corpus terms for the language pairs they use.